An Efficient Hand Gesture Recognition Framework for Sign Languages using Multiscale Attention-Based EfficientNetB7 and Optimization Techniques

In the recent era, “Hand Gesture Recognition (HGR)” is a recent research hot topic that is familiar and the most natural interaction in human–computer interaction. It is one of the most effective means of communication among normal people and hearing-impaired people. Even the differences and complications like views, illuminations, shapes, different sizes, and self-structural characteristics make HGR a challenging task. The diverse research papers are explored by analyzing the different techniques that are adapted to computer vision technology. Due to its complicated environment with limited occlusion and light effects, HG segmentation becomes a complicated task in traditional methods. Besides the success of extensive clinical trials, there are challenging phases in the powerful webcam sight translation of “Sign Language (SL),” which is one of the main applications of HGR. The effective recognition system must solve challenges such as local hand shape representation, gesture sequence modeling, hand segmentation, and global body configuration. Thus, it is important to design an innovative HGR model that depends on deep learning techniques to resolve the complications of the classical HGR model. At first, the acquired HG images are gathered from the standard datasets, providing image segmentation regions. In this phase, image segmentation is achieved using “Adaptive Deeplabv3[Formula: see text] (ADv3[Formula: see text])” and the “parameters of Deeplabv3[Formula: see text]” are tuned using the developed Modified Flight Step Size of Wild Geese Migration Optimization (MFSS-WGMO) optimization model. The segmented images were subjected to the HGR stage. Here, the Multiscale Attention-Based EfficientNetB7 (MA-ENB7) is utilized to effectively classify HGs. Thus, the recommended HGR model achieved better efficiency than the conventional models in various experimental observations. The findings of the developed model show a value of 0.94% in terms of accuracy, sensitivity, and specificity. This analysis helps the developed framework to offer better and more efficient outcomes in the HGR framework.